/

Super-resolution Pléiades/SPOT

Quadruples the perceived image resolution of Pléiades and SPOT images


Introduction

For more information, please read the block description.

Block type: PROCESSING

This block quadruples the perceived image resolution of Pléiades and SPOT satellite imagery to increase the performance of object detection algorithms.

Super-resolution is the process of increasing the resolution of images using an algorithm. The block uses a state-of-the-art Convolutional Neural Network (CNN).

Quality improvements of the images are measured using the SSIM metric, thus guaranteing that the algorithm increases the information content of the original image.

Image resolution of the processed images will be quadrupled, but it needs to be understood that an algorithmically derived image can never have the same information content as an image that was originally recorded at that resolution. The use case for this block is as a preprocessing step for object detection algorithms (ships, cars, planes, etc.) as the images become crisper and contour outlines more well defined.

Quadrupling the resolution means here that this block takes a Pléiades or SPOT image and increases the number of pixels by a factor of 16 for all existing spectral bands by using a trained CNN. From an Information theory point of view, the generated images do not contain more information than the recorded at the original resolution.

This block implements the model architectures described in [Müller2020]. The paper also describes the methodology used and resulting metrics of each model architecture.

Supported parameters

  • model: The model to use to super-resolve the image. One of SRCNN (default), AESR or RedNet.

Choosing the deeper model architectures (AESR and RedNet) will significantly impact the time required to super-resolve the image.

Example parameters using the SPOT DIMAP download block as data source, returning the super-resolved result using the AESR model:

{
  "oneatlas-spot-fullscene:1": {
    "ids": null,
    "bbox": [
      13.405215963721279,
      52.48480326228838,
      13.4388092905283,
      52.505278605259086
    ],
    "time": null,
    "limit": 1,
    "order_ids": null,
    "time_series": null
  },
  "superresolution:1": {
    "model": "AESR"
  }
}

References

Müller2020
Müller, M. U. et al. “SUPER-RESOLUTION OF MULTISPECTRAL SATELLITE IMAGES USING CONVOLUTIONAL NEURAL NETWORKS.” ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences V-1-2020 (2020): 33–40. 10.5194/isprs-annals-V-1-2020-33-2020.